Hi Bob,
Thank you very much for taking the time and the advice, especially the aprt about paying attention to the lower and upper bounds. I think that might have been problematic and led to the issues I was seeing. So the informed priors seem to be the way to go.
In the meantime I saw somewhere the concept of a soft penalisation, where basically I would penalise the target log-likelihood depending on how much a certain parameter went past a lower/upper bound. Something like this for a lower bound at a specific value:
real penalisation;
real lowerBound_log
if(meta_trans_perPat_logits[1,1, ipat] < lowerBound_log) {
target += -penalisation * (meta_trans_perPat_logits[1,1,ipat] - lowerBound_log)^2;
}
While this makes sense to me conceptually, I wouldn’t be sure of which values to use for the penalisation. In my implementation I have both meta_trans defined as a parameter, and meta_trans_perPat defined in the transformed parameters as a result of each patient’s covariate influence on meta_trans. So I suppose penalising based on meta_trans_perPat (which repeats for every patient) vs one time based on meta_trans after the patient cycle is done would imply different values.
Would definitely appreciate your thoughts on this and if it makes sense, then on what would be an effective range of values to use.
Thanks again for your time!